Vision science has identified a number of factors that affect detection threshold for spatial targets in backgrounds. Typically, simple stimuli are used to allow precise experimental control and rigorous hypothesis testing. However, an ultimate goal of vision science is to understand performance under natural conditions, where multiple factors are varying simultaneously in complex ways. We propose a direct experimental approach for identifying and quantifying the factors that affect detection performance in natural scenes. First, we obtain a large representative collection of calibrated natural images. Next, we divide the images up into millions of background patches and sort them into narrow bins along dimensions of interest. For example, in the present study each bin represents a particular (narrow range of) mean luminance, contrast, and spatial correlation of the background to a given target. Next, we measure detection thresholds in humans parametrically for a small subset of bins spanning each dimension. The psychometric function for each bin is measured by randomly sampling (without replacement) background patches from that bin. Finally, we analyze the residual variation of the background patches within each bin for other factors that strongly correlate with the measured performance. In our initial measurements with a 4-cpd Gabor target in two subjects, we find (with background-target correlation fixed) that threshold amplitude is a linear function of mean luminance (Weber's law for luminance) and threshold power is a linear function of background contrast power (Weber's law for contrast). Thus, our results suggest that these classic laws translate to natural backgrounds. However, a preliminary analysis suggests that other factors will emerge beyond the three we controlled for. Finally, we note that this general approach should be applicable to other natural tasks as long as a sufficiently large set of natural stimuli can be obtained.
展开▼
